Flies and smells

Connectomics informed modelling of olfactory learning in Drosophila Melanogaster

Rory Bedford

MRC LMB

Credit assignment

  • Deep neural networks need to set their weights effectively to function
  • Backpropagation is infeasible in biological networks
  • How networks assign credit remains a major area of research in neuroscience
  • Solutions consist of a combination of synaptic learning rules and architectural motifs

Dopaminergic error signals

  • Modulatory signals affect plasticity at the synapse
  • 3-factor learning allows weights to change in an error-dependent manner
  • Dopaminergic neurons are traditionally thought to convey global error signals
  • Zero-order optimization is inefficient in large networks
  • Research direction: study the connectivity of dopaminergic neurons in a learning centre

Fruit fly associative learning

  • Drosophila excel at classical (Pavlovian) conditioning, associating stimuli like odors with rewards or punishments
  • Learning is rapid and efficient, often requiring only a few trials
  • This provides an excellent minimal model of learning

The advent of connectomics

  • Researchers have recently constructed a complete, synaptic-resolution wiring diagram of the adult Drosophila brain (Schlegel et al. 2024)
  • Heroic work, using electron-microscopy, machine vision, and extensive curation
  • Allows us to study the learning circuitry of Drosophila in detail

Olfactory processing circuits

  • Odors are first detected by olfactory receptor neurons (ORNs) and processed in the Antennal Lobe (AL)
  • Projection Neurons (PNs) relay information to the Mushroom Body (MB), a key center for learning and memory
  • Kenyon Cells (KCs) in the MB sparsely encode odors, enhancing pattern separation
  • Dopaminergic neurons modulate synapses in response to reward or punishment, reinforcing associative learning
  • MB output neurons (MBONs) receive inputs from Kenyon Cells, and drive either attractive or aversive behaviours

Compartmentalised mushroom body structure

  • 15 compartments with distinct MBONs representing positive or negative valences
  • Kenyon cells project within a given compartment with little overlap
  • Dopaminergic neurons send error signals within their respective compartments
  • Different lobes known to have different learning rates, and are implicated in associative memory across different timescales
  • The functional role of this structure remains unknown

Research question:

Why is the mushroom body compartmentalised?

  • How do the following factors affect learning:
    • The number of compartments
    • The distribution of compartment valences
    • The distribution of compartment learning rates

The model:

  • A simplified ensemble of perceptrons
  • Output logits representing confidence in their respective valences
  • Trained individually with their own error signals
  • Compartment signals integrated to produce a global valence prediction

The task:

  • Input random binary vectors with biologically accurate dimensionality and sparsity
  • Train to predict valence scores of 1 or -1 associated with each pattern

Comparison of single- vs multi-compartments



Dynamically changing targets

Dynamically changing targets



Multiple rates of changing targets


Conclusions

  • Having a number of compartments seems to help learning somewhat
  • An ensemble of perceptrons will find an optimal learning rate for an environment with a given rate of variability in valence assignments
  • Couldn’t get a model to work with a variety of different rates of changing valences

Future work

  • The success of the ensemble model needs further theoretical and computational exploration
  • Need to find a model that works on data with a non-constant rate of change

Thank you